A critical barrier to the trustworthiness of sixth-generation (6G) agentic autonomous networks is the uncertainty neglect bias; a cognitive tendency for large language model (LLM)-powered agents to make high-stakes decisions based on simple averages while ignoring the tail risk of extreme events. This paper proposes an unbiased, risk-aware framework for agentic negotiation, designed to ensure robust resource allocation in 6G network slicing. Specifically, agents leverage Digital Twins (DTs) to predict full latency distributions, which are then evaluated using a formal framework from extreme value theory, namely, Conditional Value-at-Risk (CVaR). This approach fundamentally shifts the agent's objective from reasoning over the mean to reasoning over the tail, thereby building a statistically-grounded buffer against worst-case outcomes. Furthermore, our framework ensures full uncertainty awareness by requiring agents to quantify epistemic uncertainty -- confidence in their own DTs predictions -- and propagate this meta-verification to make robust decisions, preventing them from acting on unreliable data. We validate this framework in a 6G inter-slice negotiation use-case between an eMBB and a URLLC agent across 200 trials. The results demonstrate the profound failure of the biased, mean-based baseline, which systematically violates the strict URLLC SLA 11 times. Our unbiased, CVaR-aware agent successfully mitigates this bias, eliminating SLA violations entirely and significantly reducing the 99.999th-percentile latencies by up to 51.7\%. We show this reliability comes at the rational and quantifiable cost of reduced energy savings, exposing the false economy of the biased approach. Crucially, executing our framework with an otel-llm-1b-it model on a single NVIDIA RTX A4000 GPU achieves sub-1.5-second inference times, validating the feasibility for non-real-time RIC use-cases.

翻译：第六代（6G）自主智能网络可信度的关键障碍在于不确定性忽视偏差——即大语言模型（LLM）驱动的智能体倾向于基于简单平均值做出高风险决策，而忽略极端事件的尾部风险。本文提出一种无偏、风险感知的智能协商框架，旨在确保6G网络切片中鲁棒的资源分配。具体而言，智能体利用数字孪生（DT）预测完整时延分布，并采用极值理论中的正式框架——条件风险价值（CVaR）对其进行评估。该方法将智能体的优化目标从均值推理根本性地转向尾部推理，从而构建基于统计的最差情形缓冲机制。此外，我们的框架通过要求智能体量化认知不确定性（即对自身DT预测的置信度）并传播这种元验证以做出稳健决策，确保完全的不确定性感知，防止基于不可靠数据行事。我们在eMBB与URLLC智能体之间的6G跨切片协商场景中，通过200次试验验证了该框架。结果表明，存在偏差的均值基线方法存在严重缺陷，系统性地违反严苛的URLLC服务等级协议（SLA）达11次。而我们的无偏CVaR智能体成功缓解了这种偏差，完全消除了SLA违规，并将99.999百分位时延显著降低达51.7%。我们证明这种可靠性以合理且可量化的能源节约损失为代价，暴露了偏差方法的虚假经济性。关键的是，在单个NVIDIA RTX A4000 GPU上使用otel-llm-1b-it模型执行我们的框架，推理时间低于1.5秒，验证了其在非实时无线接入网智能控制器（RIC）用例中的可行性。